Your Mac hears everything. Now you can read it.
당신의 맥은 모든 것을 듣고 있어요. 이제 읽을 수 있습니다.
→ ko · 한국어 · one target language per session
pildam live-captions and live-translates any app's audio — and your own voice — entirely on this machine. Nothing you hear or say ever leaves it.
You sit in meetings, lectures, and calls where more than one language is in the room. pildam puts a quiet glass caption card over whatever you're hearing — any app's audio, your own microphone, or both at once with echo cancellation — and translates it as it's said. It lives in the menu bar, starts when you ask, and writes everything into a record you can actually use later.
No account. No analytics. No telemetry. Nothing leaves the machine.
I.It hears everything
System audio from every app, or only the apps you pick — calls, videos, lectures, meetings. Your microphone runs alongside with voice-processing echo cancellation, so a meeting and your own voice are captioned together without double-captions.
a speaker chip — streaming diarization, live
b dimmed tail — a live partial, still being recognized
c gray line — tentative translation, trailing ~1.5 s; the final decodes over it token by token
When nothing appears, pildam says why.
Apple's Live Captions go silently blank. pildam watches the audio on a one-second heartbeat and names the problem instead:
“No audio detected — check that the source you're capturing is actually playing.” The status is a sentence, never a spinner.
And one toggle keeps the overlay out of screenshots, recordings, and shared screens — only you see the captions.
II.It translates like an interpreter
We could say instant. It isn't. It's 1.5 seconds. While someone speaks, a tentative translation trails the live words — the way a human interpreter trails a speaker — and the final translation decodes token by token into the caption. Long monologues are cut at breath and clause lulls (8 s soft, 14 s hard) so the translator always gets whole clauses.
spoken words above the line · translation below · same clock
- Three latency tiers
- 560 / 1120 / 2240 ms audio chunks — caption sooner, or transcribe more accurately.
- Pin a language, or don't
- Pinning improves accuracy when you know what will be spoken. A Strict Language Lock exists too — and the app itself tells you to leave it off when speech mixes languages, like English terms in Korean.
- One target per session
- Everything is translated into the language you chose — 24 to pick from. It doesn't translate arbitrary pairs simultaneously.
III.It never phones home
Privacy here is architecture, not policy. The entire pipeline runs on your Mac's own silicon:
Silero VAD v61 → Nemotron-3.5 ASR 0.6 B2 → LS-EEND diarization3 → Hy-MT2 1.8 B · 1.25-bit4
1 voice activity, 256 ms streaming — CoreML
2 multilingual streaming speech recognition — on the Neural Engine, ~17× realtime headroom on an M2 Pro
3 who is speaking, live — CoreML
4 translation on the GPU via MLX, ~144 tokens/s on an M2 Pro — llama.cpp CPU fallback included
| # | request | when | size |
|---|---|---|---|
| 1 | Model download, from Hugging Face | once, on first run | ~2.4 GB |
| 2 | Update check — cryptographically signed, always asks before installing | daily · can be turned off | a few KB |
After that, you could unplug the router.
IV.The record
Everything you caption lands in History — a transcript archive that lives only on this Mac. Verbatim is never thrown away: Polish rewrites the sentence; the record keeps what was actually said.
Sessions grouped by day, folders and favorites, find-in-transcript. Your full history stays readable without paying anything, ever.
The session's audio with a speaker-colored waveform, karaoke word highlighting, click-a-line-to-seek, 0.5–2× speed.
Merges each speaker's streamed fragments into clean sentences and re-translates the whole thing — verbatim kept behind a toggle.
Right-click a caption, teach it the right term — churn → 고객 이탈. The glossary applies to live captions and polish alike.
A title, key points, and an abstract per recording — Apple's on-device model, so it needs Apple Intelligence and uploads nothing.
SRT (dual-language cues if you want), WebVTT, plain text, Markdown, JSON — save panel or share sheet.
V.The fine print, set large
- macOS 26 on Apple Silicon. No Intel builds. We'd rather tell you here than in a refund email.
- A one-time ~2.4 GB model download before anything works. Stored only on your Mac. Then fully offline.
- It asks to hear your microphone and system audio. You're responsible for recording consent where the law requires it.
- Near-real-time, not instant. The fastest tier captions in 560 ms; translation trails by ~1.5 s. That's the honest shape of interpreting.
- Not on the App Store. A notarized Developer ID download. Updates are signed, checked daily, and always ask first.
- Version 0.9. Transcripts and translations are machine-generated and can be wrong. Verify before you rely on them.
VI.Price, once
Free
$0 · forever
The full live engine, uncompromised — captions, translation, all 24 languages, and your entire history stays readable. Not a trial.
Pro
$39 · one-time
The workflow on top: export, playback, summaries, polish, glossary. Up to 2 Macs, personal use.
Business
$79 · one-time
Everything in Pro, licensed for commercial use and teams. Up to 3 Macs, priority support.